CN103198605A - Indoor emergent abnormal event alarm system - Google Patents

Abstract

The invention discloses an indoor emergent abnormal event alarm system. The indoor emergent abnormal event alarm system comprises a video collecting module, an audio collecting module, a heat releasing infrared detection module, a vibration detection module, a glass breaking detection module, a door magnet detection module, a signal processing module, a wireless receiving module, an alarm sending module, an alarm linkage module, a voice output module, an infrared light-emitting diode (LED) light-compensating lamp module, and a power management module. The indoor emergent abnormal event alarm system conducts real-time dynamic detection analysis on emergent abnormal events that people fall down and lay on the floor for a long time, illegal break-in, theft, violent physical confrontation, fire accidents, coal gas leakage, explosion and the like in an indoor deployed and controlled area, sends intuitive alarm signals which take audios, videos or pictures as carriers in a wire communication mode or a wireless communication mode to alarm receiving and processing terminals such as a mobile phone or a computer of a householder and residential area monitoring computers in the first time, associates with the audible and visual alarm device to give an alarm, and receives voice output of receiving terminal devices to achieve voice talkback.

Description

Indoor unexpected abnormality affair alarm system

Technical field

The present invention relates to a kind of warning system, be specifically related to a kind of indoor unexpected abnormality affair alarm system.

Background technology

We know, in order to avoid theft, generally can be in the residential quarter or outdoor corridor, door and window etc. locate to install the anti-theft device of camera and so on, a kind of monitoring and warning system of swarming into based on the indoor occupant of intelligent video for example, mainly be by the intellectual analysis to camera image, realize the analysis to indoor place moving target, breaking in of personnel reached timely discovery and warning, ignore the motion that is safe from danger simultaneously, reduce false-alarm.Also have a kind of home furnishings intelligent burglary-resisting system, it is made of antitheft door, mechanical lockset, window, terrace and balcony door at least.By being installed in: the door net sensor on (1) antitheft door, antitheft door Magnetic Sensor, (2) be installed in dead bolt sensor, key on the lockset and insert/extract sensor, the list sensor of beating, (3) be installed in the interior handling position sensor of outer handle subassembly, arrange simultaneously from terrace and enter balcony path detecting area the balcony door passage, enter window path detecting area the passage in room from window, realize owner in the identification at home automatically and robber's Smart Home burglary-resisting system.

There is following shortcoming in existing door and window anti-intrusion system:

1. the most product function in existing market is single, and does not possess comprehensive analytical capacity.

2. there is defective separately in single anti-Intrusion Detection Technique, in some cases can't operate as normal, in case can't operate as normal, whole detection system is in paralyzed state, the poor stability of system.As: intelligent video analysis detects the interference that may be subjected to irrelevant motions such as ambient light variation, minute surface reflection, causes a large amount of wrong reports; Rpyroelectric infrared detects the interference that is subject to temperature, high light, circumstance complication motion; Vibration detection is vulnerable to the interference that external force such as wind causes that the door and window vibration produces; The broken detection of glass is vulnerable to the interference of external environment noise; Door magnetic detects needs door and window strict closed, can't use under the indoor situation that needs to ventilate, and as the summer of sweltering heat, the user often needs the ventilation of windowing.

3. traditional warning system relates to a large amount of various sensors as " home furnishings intelligent burglary-resisting system ", and Installation and Debugging are very complicated usually, and warning message is not directly perceived simultaneously, can't learn field condition;

4. traditional warning system can't be worked under the situation of cutting off external power supply, gives illegal invasion person with opportunity.

5. can't carry out certain monitoring to indoor competent person, for example children are often gone out or solitary old man, its health status is to need certain measure to be monitored, and can in time save as falling etc. when guaranteeing abnormal conditions to occur, and does not have this class function at present.

Summary of the invention

The object of the present invention is to provide a kind of indoor unexpected abnormality affair alarm system, solved existing warning system function singleness, rate of false alarm height, easy disturbed paralysis can't operate as normal, and can't fall, suffer the problem of insurgent violence abnormal event alarming to competent person such as old man.

For solving above-mentioned technical matters, the present invention by the following technical solutions:

A kind of indoor unexpected abnormality affair alarm system is characterized in that: comprise

Video acquisition module adopts video sensor to gather video stream signal, finishes the digitizing of picture signal, and the pre-service of picture signal is met the digital video signal of signal processing module requirement and exports to signal processing module;

The audio collection module is finished the digital-to-analog conversion of voice signal, and the sample code of voice signal and filtering are handled, and is met the digital audio and video signals of signal processing module requirement and exports to signal processing module;

Heat discharges the infrared detection module, the difference of the temperature of inducing moving objects and background object, and heat is released the infrared different information that can sense human body temperature and ambient temperature when human body moves, and converts the output of voltage signal backward signal processing module to;

The vibration detection module produces extraneous vibration deformation or is subjected to force information to change voltage signal into, exports to signal processing module;

The broken detection module of glass and door magnetic detection module are transformed into voltage signal by the wired or wireless communication mode with corresponding information and export to signal processing module;

Signal processing module, collect heat and discharge the signal that infrared detection alarm module, vibration detection module, video acquisition module and audio collection module transmit, carry out analysis-by-synthesis, differentiate illegal invasion, violent conflict, fall do not rise for a long time, anomalous event such as gas leak, and send alerting signal to the warning sending module;

Wireless receiving module is used for receiving the vibration detection module, the broken detection module of glass, the alerting signal that the alarm sensor on smoke detector, emergency call button and the door magnetic detection module sends;

The warning sending module adopts the wired or wireless communication mode, is mainly used in receiving alerting signal that signal processing module sends and to sending warning message to householder's mobile phone or cell management center;

The alarm linkage module, interlock sound and light alarm equipment is worked simultaneously;

The voice output module realizes the voice output from RTU (remote terminal unit) such as mobile phone, Surveillance center;

The infrared LED light supplementation lamp module when ambient light illumination is not enough, provides secondary light source to video acquisition module, guarantees that it still can collect effective view data under the environment of low-light (level);

Power management module connects external power supply and accumulator, when the external power supply normal power supply, entire product is used external power supply, when extraneous power cut-off, inner standby battery is enabled automatically, guarantees that above-mentioned each module continues operate as normal after external power interruption.

Further technical scheme is that product is integrated in the casing, described video acquisition module, audio collection module, heat discharge infrared detection alarm module, vibration detection module, warning sending module, wireless receiving module, alarm linkage module and warning sending module and are installed in the casing, described signal processing module and power management module and switch lamp control module are installed in the shell middle part, and the infrared LED light supplementation lamp module is arranged between product outward flange and the middle part.

Further technical scheme is that kernel processor chip adopts the double-core architecture mode in the above-mentioned signal processing module, namely " primary processor+from processor " the double-core architecture mode, primary processor is mainly finished the collection of audio-video signal, audio/video coding, the rpyroelectric infrared alerting signal is collected, receive the vibration detection module by wireless receiving module, the broken detection module of glass, door magnetic detection module, the alerting signal that smoke detector, emergency call button distributed alarm sensor send sends the multi-source alerting signal from processor to together with audio, video data; Be mainly used in moving the audio-video intelligent analysis algorithm from processor, judge whether to exist the unexpected abnormality event by video and sound, simultaneously in conjunction with the rpyroelectric infrared alerting signal, the vibration alarming signal, the broken alerting signal of glass, multi-source informations such as door magnetic alerting signal, form final decision signal, take place if determine anomalous event, send alerting signal to the primary processor end, and start the audible-visual annunciator warning.

Further technical scheme is that above-mentioned householder's identity is judged and comprised by recognition of face and carry out identity validation that wherein said recognition of face mainly is divided into

The front face photo of same individual's different angles is gathered in the registration of people's face, finds the position of people's face in image by people's face location, and standardization people face size, people's face angle and illumination, the feature of extraction standardization facial image, typing registered user database;

Recognition of face, adopt the HAAR feature to realize that in conjunction with the adaboost algorithm people's face detects realization people face location, adopted two-layer human eye steady arm, all be to obtain the human eye location by the adaboost algorithm, utilize people's face positioning result by the image rotation eyes to be proofreaied and correct and be the standardization of level realization people face, carry out feature extraction by two-dimensional Gabor filter, utilize Euclidean distance between vector as coupling tolerance mode, realize the facial image of inquiry and the coupling of database facial image by the arest neighbors classification.

Further technical scheme is above-mentioned signal processing module

Be connected with memory module;

Be connected with external power supply and battery by power management module;

Be connected with the internal memory for the master cpu operation;

Be connected with the flash memory for storage system start-up routine, configuration parameter, log information;

Be connected with by external warning lamp, the warning signal interlocking equipment is realized local sound and light alarm, or controls the interlink warning module of other associate device.

Further technical scheme is that above-mentioned video acquisition module comprises common shooting detection module and the degree of depth shooting detection module that view data is alignd mutually, obtains sextuple information component＜x, y for 1 P in the image, d, r, g, b＞, wherein＜and x, y＞be the coordinate of picture element in image, d is that shot object is apart from the distance of video camera,＜r, g, b＞be color component, this indoor unexpected abnormality affair alarm system for detection of moving target and tracking method as follows:

Coloured image is converted to gray level image, sets up background model at gray level image and depth image respectively, adopt mixed Gaussian to set up background model, the probability distribution that mixed Gaussian background modeling hypothesis signal changes can be used K Gaussian distribution match, is expressed as

$\mathrm{prob}\left(x\right)=\underset{i=1}{\overset{K}{\mathrm{\Σ}}}{\mathrm{\ω}}_i{\mathrm{\η}}_i(x,{\mathrm{\μ}}_i,{\mathrm{\σ}}_i)$

Wherein μ and σ are average and the variance of Gaussian distribution, each Gaussian distribution η in the model _i(x, μ _i, σ _i) all give a weights omega _i, μ wherein _iAnd σ _iAverage and the standard deviation of representing Gaussian distribution respectively, a plurality of Gaussian distribution obtain the probability distribution of signal by linear combination, K Gaussian distribution is according to the descending sort of ω/σ, arrange forward Gaussian distribution and enough represent the distribution of background, mixed Gauss model can be safeguarded the variation of scene automatically simultaneously, situation about surveying for flase drop can be righted the wrong by study simultaneously

With preceding B Gaussian distribution model as a setting, remaining Gaussian distribution is thought prospect, and B should satisfy

$B=\arg \mathrm{mi}{n}_b(\underset{n=1}{\overset{b}{\mathrm{\Σ}}}{\mathrm{\ω}}_n\left(x\right)>T)$

Present picture element value I (x) and a preceding B Gaussian distribution are mated, if the match is successful then this pixel is background pixels with wherein any one Gaussian distribution, otherwise are the foreground moving pixel, matching way as shown in the formula

|I(x)-μ _i(x)|＜c×σ _i(x)i＝1.....B；

Goal verification, mode by background modeling detects and has obtained moving object, but a lot of motions all are to be changed by indoor light, the swing of curtain, the generation of motion such as the flicker of television image, and the monitored object here is the people, therefore people's motion need be separated from a large amount of irrelevant motions.Here by the light interference filter, people's face detects filtrator and a shoulder detects filtrator, and it is main adopting the gray level image analysis, and the depth image analysis is auxilliary, detects the affirmation that realizes monitoring objective in conjunction with the detection of people's face and head shoulder simultaneously,

Light interference filter wherein: the variation of light makes the gray-scale value of image that variation take place, the accommodation that has surpassed background model when speed and the amplitude of gray-value variation, then detect and be moving target, filter light by depth image and change the motion that causes, depth image is owing to adopted the infrared light detecting method, substantially be not subjected to the interference of illumination variation, detect by the gray level image background modeling and obtain the moving region and be

, the moving region in the corresponding depth image is

, following condition must be satisfied in real target area:

$\frac{\underset{x\∈R_d^i}{\mathrm{\Σ}}{P}_d\left(x\right)}{\underset{x\∈R_g^i}{\mathrm{\Σ}}{P}_g\left(x\right)}>T$

P wherein _d(x) and P _g(x) be respectively depth image and gray level image motion detection result, P (x)=1 represents motion pixel, and P (x)=0 represents background pixels;

People's face and head shoulder detect filtrator: have a large amount of real motions in the actual monitored scene, rotation as fan, the swing of curtain, the flicker of television image, the motion of pet, and the present invention only is concerned about people's motion, the people has the obvious external appearance characteristic that is different from these motions, as face characteristic and head shoulder feature, here adopt the haar small echo to realize that in conjunction with the adaboost sorter people's face detects in gray level image, the HOG feature realizes that in conjunction with the svm classifier device head shoulder detects, and is illustrated as the people of motion when detecting face characteristic or head shoulder feature in the moving region, otherwise think to disturb, give filtering;

Target following by the target association of interframe, forms the movement locus of target, for the succeeding target behavioural analysis provides foundation on the basis of goal verification.Target following mainly comprises: target prodiction, and target signature is selected, and target association coupling and target signature are upgraded, wherein

Target prodiction be according to present frame and before the location estimation target of target in the position of next frame, be conducive to improve the precision of subsequent association coupling.Here adopt the mode of estimating target speed, utilize following formula predicted position,

$v_t^x=\frac{\underset{n=0}{\overset{N-2}{\mathrm{\Σ}}}(x_{t-n}-x_{t-n-1})}{N-1}$ With $v_t^y=\frac{\underset{n=0}{\overset{N-2}{\mathrm{\Σ}}}(y_{t-n}-y_{t-n-1})}{N-1}$

Wherein

With

Target is in the speed of x direction and y direction constantly to represent t respectively, and N is time window, x _T-nAnd y _T-nRepresent t-n horizontal ordinate and the ordinate at the extraneous rectangle frame of target center, in like manner x constantly respectively _T-n-1And y _T-n-1Represent t-n-1 horizontal ordinate and the ordinate at the extraneous rectangle frame of target center constantly respectively.；

It is two features of center C of selecting relatively stable reliable target area S and target boundary rectangle frame that target signature is selected.The target association coupling is to find the former frame target in the position of present frame, realize the location of target, according to the correlation tracking principle, as long as guarantee under the situation of enough image sampling rates, the change in location of same target between adjacent two frames is not too large, therefore the hunting zone of target can be limited to a less distance range, the while also greatly reduces the risk of mistake coupling, establishes t moment target j to be

Target i is constantly

When satisfying following formula, just carry out characteristic matching:

$\mathrm{dist}\left(O_t^j\right\{C\},O_{t-1}^i\{C\left\}\right)<\mathrm{\&gamma;}$

Wherein γ is the detection range threshold value, need determine according to actual scene.

The criterion of coupling is following formula

$\arg \min (\mathrm{\&omega;}\&times;\left(\frac{\mathrm{dist}\left(O_t^j\right\{C\},O_{t-1}^i\{C\left\}\right)}{\mathrm{\&gamma;}}\right)+(1-\mathrm{\&omega;})\&times;\frac{\left|O_t^j\right\{S\}-O_{t-1}^i\{S\left\}\right|}{O_{t-1}^i\left\{S\right\}})<T$

Wherein ω is the weight factor of feature, value 0.5, and T is the error upper limit, prevents the mistake coupling, value 0.4.

The data of filtering sudden change are all adopted in the renewal of target signature, guarantee that the single order smooth mode renewal of bigger fluctuation can not appear in data, and specific implementation is the formula of delegation as follows

$O_t^i\left\{F\right\}=\mathrm{\&alpha;}\&times;O_{t-1}^i\left\{F\right\}+(1-\mathrm{\&alpha;})\&times;O_t^i\left\{F\right\},F=\{S,C\}$

Wherein α is for upgrading the factor, value 0.2.

The indoor unexpected abnormality affair alarm of basis system is as follows for detection of the method for falling, the automatic three-dimensional modeling in ground, the video camera that will link to each other with video acquisition module is tilted to down and the angled installation in ground, the initial point of camera coordinate system is set on the Z axle of image coordinate system, and translation matrix T is reduced to like this $\left[\begin{array}{c}0\\ 0\\ H\end{array}\right],$ Wherein H represents the video camera photocentre to the distance on ground, and the conversion formula of camera coordinate system and world coordinate system is as follows

$\left[\begin{array}{c}x\\ y\\ z\\ 1\end{array}\right]=\begin{array}{c}\left[\begin{array}{cc}R& T\\ 0^T& 1\end{array}\right]\end{array}\left[\begin{array}{c}{x}_w\\ y_w\\ z_w\\ 1\end{array}\right],T=\left[\begin{array}{c}0\\ 0\\ H\end{array}\right],R=R_x\left(\mathrm{\&alpha;}\right)\&CenterDot;R_y\left(\mathrm{\&beta;}\right){\&CenterDot;R}_z\left(\mathrm{\&gamma;}\right),$

Three coordinate axis of camera coordinate system and three coordinate axis angles of world coordinate system are α, and β and γ suppose that the initial point of image coordinate system overlaps with the initial point of camera coordinate system, then in the depth image a bit (d) coordinate under camera coordinate system is for u, v

Wherein (u v) is the coordinate of picture element in image, and d is that shot object is apart from the distance of video camera, f _xAnd f _yFocal length for video camera level and vertical direction on imaging plane;

If ground some F (x, y z) satisfy following plane restriction under camera coordinate system:

ax+by+cz+d＝0

Wherein

Be the normal vector of ground level, can calculate the angle α of three coordinate axis, β and γ by normal vector:

$\mathrm{\&alpha;}=\arccos \frac{a}{\left|\stackrel{\&RightArrow;}{v}\right|},$ $\mathrm{\&beta;}=\arccos \frac{b}{\left|\stackrel{\&RightArrow;}{v}\right|},$ $\mathrm{\&gamma;}=\arccos \frac{c}{\left|\stackrel{\&RightArrow;}{v}\right|}$

After obtaining frame depth image data, adopt the RANSAC algorithm to realize plane fitting, can obtain a plurality of candidates' ground level F by match _I=1...n{ a _i, b _i, c _i, d _i, by following priori coarse sizing is carried out on the plane; : one, ground has occupied bigger area in image, and namely real ground level should comprise more image pixel point; Two, generally the angle γ of video camera and z axle between 40 degree and 80 degree, and the angle α of x axle at 0 degree between 20 degree, equal 0 degree substantially with the angle β of y axle.

Select in the remaining plane behind the coarse sizing from camera plane farthest namely to satisfy as ground level:

$F=\underset{1\&le;i\&le;n}{\arg \max }{d}_i$

Any point to the distance of ground level is under the camera coordinates:

$h=\frac{|\mathrm{ax}+\mathrm{by}+\mathrm{cz}+d|}{\sqrt{a^2+b^2+c^2}}$

Calculate target's center to the distance of ground level by following formula;

The static judgement of target, judge by the time-domain difference that calculates gray level image in the target area whether target is static:

$M=\frac{\underset{(x,y)\&Element;R}{\mathrm{\&Sigma;}}|g_t(x,y)-g_{t-1}(x,y)|}{S_R}$

G wherein _t(x, y) constantly in the gray level image (x y) locates the gray-scale value of picture element to expression t, and R represents the target area, S _RElemental area for the target area.The explanation target is static when M＜ε, and wherein ε is a very little positive number.

Further technical scheme is that the audio collection module also detects specific sound, and its detection method bag is as follows:

The voice signal pre-service, the sampling rate of supposing sound signal X (t) is f _s, f _sValue 8kHz passes through pre-emphasis successively with X (t), divides frame and windowing process, and window function is selected Hanning window, and removes average, avoids DC component that near the spectral line ω=0 place is exerted an influence;

The period map method in the Classical Spectrum estimation is adopted in feature extraction, uses FFT to realize, finally obtains normalized power spectrum X (f _n), the extraction number is 24 Mel bank of filters.Power spectrum X (f _n) through the Mel bank of filters filtering take the logarithm, obtain Mel cepstrum MFCC coefficient through discrete cosine transform again, the Mel bank of filters is made up of one group of V-belt bandpass filter that distributes according to the Mel frequency marking;

The training of GMM audio frequency identification model adopts the maximum EM algorithm of expectation to ask for the GMM model, given training sample set X={x ₁, x ₂..., x _n, the likelihood function of GMM is

$p(X/\mathrm{\&lambda;})=\underset{i=1}{\overset{n}{\mathrm{\&Pi;}}}p(x_i/\mathrm{\&lambda;})$

Model parameter wherein p _iThe probability of expression Gauss model,

Σ _iMean vector and the covariance matrix of representing Gauss model respectively.

The EM algorithm comprised for two steps, and the E step is asked for expectation, calculates auxiliary function

M goes on foot expectation maximization, maximization

Obtain The continuous iteration that goes on foot by E step and M is until algorithm convergence,

$Q(\mathrm{\&lambda;},\widehat{\mathrm{\&lambda;}})=\underset{y}{\mathrm{\&Sigma;}}\left(P(Y=y|X=x\left|\mathrm{\&lambda;}\right)\log P(Y=y,X=x|\widehat{\mathrm{\&lambda;}})\right)$

Wherein X is observed reading, and Y is implicit state.

When the expectation value maximal value of adjacent twice iterative computation was more or less the same, then algorithm convergence stopped iteration, is shown below:

$Q_t(\mathrm{\&lambda;},\widehat{\mathrm{\&lambda;}})-Q_t(\mathrm{\&lambda;},\widehat{\mathrm{\&lambda;}})<\mathrm{\&epsiv;}$

Wherein t represents number of iterations, and ε is a less positive number;

The identification of use training pattern obtains model parameter λ by the GMM training, sends in the GMM model after the sound bite extraction feature and calculates similarity, judges the classification of this sound bite by similarity.

Further technical scheme is that the method that detects the limbs conflict comprises light stream vector analysis and audio analysis, when the conflict of outburst limbs, be accompanied by violent random motion and loud uttering long and high-pitched sounds, therefore can detect the limbs conflict by light stream vector analysis and audio analysis, when two kinds of methods all detect the limbs conflict, trigger the limbs conflict and report to the police, capture a scene photograph simultaneously.

Light stream vector is analyzed, and can obtain the zone at target place by target following, adopts the light stream vector V={ ν in the KLT unique point optical flow computation target area ₁, ν ₂..., ν _n, adopt amplitude weighting histogram H _p={ h _j} _{J=1,2 ..., n}Realize the statistical study of area light flow vector, draw j rank Nogata h by following formula again _j,

$h_j=C_h\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}{A}_{v_i}\mathrm{\&delta;}(b\left(v_i\right)-j)$

Here exponent number can value 12, C _hBe normalized parameter,

Be the normalization light stream vector

Amplitude, b (v _i) be light stream vector v _iCorresponding histogram determines that by the direction of vector δ (.) is Kronecker delta function,

Adopt regional entropy E _HRealize the tolerance of violent random motion, E _HExpression formula as follows:

$E_H=-\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{h}_j\log h_j$

H wherein _jRepresent j rank amplitude weighting histogram.E _HMotion Shaoxing opera in the more big declare area is strong random, and setting threshold T works as E _HBroken out the limbs conflict during＞T in the declare area.

Compared with prior art, the invention has the beneficial effects as follows: the present invention is with the goal behavior analysis, recognition of face, intelligent audio frequency and video analysis such as voice recognition is core technology, and in conjunction with advanced rpyroelectric infrared detection, vibration detection, audio detection, glass is broken to be detected, Internet of Things and wireless communication technologys such as the detection of door magnetic and Smoke Detection, the interior personnel in zone that deploy to ensure effective monitoring and control of illegal activities in the Real-time and Dynamic Detection analysis room fall and do not rise for a long time, break in, steal, serious limbs conflict, fire, gas leak, blast waits the paroxysmal abnormality security incident, and be that the alerting signal directly perceived of carrier is by the wired or wireless communication mode with audio frequency and video or picture, the very first time is sent to such as householder's mobile phone or computing machine, cell monitoring computing machines etc. are reported to the police and are received processing terminal, the audible-visual annunciator that links is simultaneously reported to the police, and the voice output of receiving terminal apparatus, realize speech talkback; The present invention has also adopted the false-alarm filtrator of checking algorithm, moving target signature analysis based on anomalous event, merge the multi-source detection information, solved safety-protection system (as infrared eye, intelligent anti-theft lock etc.) rate of false alarm height in the conventional chamber preferably, problem such as reliability is relatively poor; At last, this product carries standby battery, can be for self powering 2-24 hour under the situation of no external power supply.

Description of drawings

Fig. 1 is the connection diagram of the indoor unexpected abnormality affair alarm of a present invention embodiment of system.

Fig. 2 is the hardware connection diagram of the indoor unexpected abnormality affair alarm of the present invention system.

Fig. 3 is the software architecture diagram of the bright indoor unexpected abnormality affair alarm of the present invention system.

Fig. 4 is primary processor end schematic flow sheet in the signal processing module of the indoor unexpected abnormality affair alarm of the present invention system.

Fig. 5 is from processor end schematic flow sheet in the signal processing module of the indoor unexpected abnormality affair alarm of the present invention system.

Fig. 6 is the registration of people's face and the recognition of face schematic flow sheet of the indoor unexpected abnormality affair alarm of the present invention system.

Fig. 7 analyzes synoptic diagram to accessing facial image during for people's face standardization of the indoor unexpected abnormality affair alarm of the present invention system.

Fig. 8 is used for the fall detection schematic flow sheet for the indoor unexpected abnormality affair alarm of the present invention system.

Fig. 9 for the indoor unexpected abnormality affair alarm of the present invention system be used for fall detection the time coordinate modeling synoptic diagram.

Figure 10 is the specific sound detection schematic flow sheet of the indoor unexpected abnormality affair alarm of the present invention system.

Figure 11 is the product experimental system connection diagram of the indoor unexpected abnormality affair alarm of the present invention system.

Embodiment

In order to make purpose of the present invention, technical scheme and advantage clearer, below in conjunction with drawings and Examples, the present invention is further elaborated.Should be appreciated that specific embodiment described herein only in order to explaining the present invention, and be not used in restriction the present invention.

Fig. 1 shows an embodiment of the indoor unexpected abnormality affair alarm of the present invention system: a kind of indoor unexpected abnormality affair alarm system comprises

Video acquisition module adopts video sensor to gather video stream signal, finishes the digitizing of picture signal, and the pre-service of picture signal is met the digital video signal of signal processing module requirement and exports to signal processing module;

The audio collection module is finished the digital-to-analog conversion of voice signal, and the sample code of voice signal and filtering are handled, and is met the digital audio and video signals of signal processing module requirement and exports to signal processing module;

Heat discharges the infrared detection module, the difference of the temperature of inducing moving objects and background object, and heat is released the infrared different information that can sense human body temperature and ambient temperature when human body moves, and converts the output of voltage signal backward signal processing module to;

The vibration detection module produces extraneous vibration deformation or is subjected to force information to change voltage signal into, exports to signal processing module;

The broken detection module of glass and door magnetic detection module are transformed into voltage signal by the wired or wireless communication mode with corresponding information and export to signal processing module, and the auxiliary signal processing module is made corresponding judgement;

Signal processing module, collect heat and discharge the signal that infrared detection alarm module, vibration detection module, video acquisition module and audio collection module transmit, carry out analysis-by-synthesis, differentiate illegal invasion, violent conflict, fall do not rise for a long time, anomalous event such as gas leak, and send alerting signal to the warning sending module;

Wireless receiving module is used for receiving the vibration detection module, the broken detection module of glass, the alerting signal that the alarm sensor on smoke detector, emergency call button and the door magnetic detection module sends;

The warning sending module adopts the wired or wireless communication mode, is mainly used in receiving alerting signal that signal processing module sends and to sending warning message to householder's mobile phone or cell management center;

The alarm linkage module, interlock sound and light alarm equipment is worked simultaneously;

The voice output module realizes the voice output from RTU (remote terminal unit) such as mobile phone, Surveillance center;

The infrared LED light supplementation lamp module when ambient light illumination is not enough, provides secondary light source to video acquisition module, guarantees that it still can collect effective view data under the environment of low-light (level);

Power management module connects external power supply and accumulator, when the external power supply normal power supply, entire product is used external power supply, when extraneous power cut-off, inner standby battery is enabled automatically, guarantees that above-mentioned each module continues operate as normal after external power interruption.

Fig. 1 also shows a preferred embodiment of the indoor unexpected abnormality affair alarm of the present invention system, this system is installed in the casing, described video acquisition module, audio collection module, heat discharge infrared detection alarm module, vibration detection module, warning sending module, wireless receiving module, alarm linkage module and warning sending module and are installed in the casing, described signal processing module and power management module and switch lamp control module are installed in the shell middle part, and the infrared LED light supplementation lamp module is arranged between casing outward flange and the middle part.

Fig. 2 shows another embodiment of the indoor unexpected abnormality affair alarm of the present invention system, kernel processor chip adopts the double-core architecture mode in the signal processing module, namely " primary processor+from processor " the double-core architecture mode, primary processor is mainly finished the collection of audio-video signal, audio/video coding, the rpyroelectric infrared alerting signal is collected, receive the vibration detection module by wireless receiving module, the broken detection module of glass, door magnetic detection module, smoke detector, the alerting signal that emergency call button distributed alarm sensor sends sends the multi-source alerting signal from processor to together with audio, video data; Be mainly used in moving the audio-video intelligent analysis algorithm from processor, judge whether to exist the unexpected abnormality event by video and sound, simultaneously in conjunction with the rpyroelectric infrared alerting signal, the vibration alarming signal, the broken alerting signal of glass, multi-source informations such as door magnetic alerting signal, form final decision signal, take place if determine anomalous event, send alerting signal to the primary processor end, and start the audible-visual annunciator warning.

As shown in Figure 3, the software of the indoor unexpected abnormality affair alarm of the present invention system mainly is made up of main treatment progress and Intelligent treatment process two large divisions, main treatment progress is made of signal acquisition module, voice output module and warning sending module, operates in the ARM end; The Intelligent treatment process is made up of intelligent analysis module and Multi-source Information Fusion false-alarm filtering module, operates in the DSP end.Signals collecting has comprised the audio-video collection module, rpyroelectric infrared alerting signal acquisition module and wireless alarm signal acquisition module, collection and the voice output of signals such as responsible audio frequency and video.For the various signals that guarantee to gather can be synchronous, so that subsequent treatment is done suitable time-delay to various alerting signals and handled with the audio-video collection signal synchronous.The Intelligent treatment process is finished intelligent audio frequency and video analysis, and Multi-source Information Fusion and false-alarm are filtered, and sends alerting signal to main treatment progress, and main treatment progress sends warning message or starts the interlink warning equipment alarm to exterior terminal by the warning sending module.

As shown in Figure 4, main process software flow process: after system powers on, at first finish the initialization of primary processor end program, next finish peripheral collecting device, the communication facilities initialization starts from processor end program.Create audio-video collection thread, rpyroelectric infrared detection thread, wireless receiving thread, send thread from processor end communication thread, warning.

Audio-video collection thread: combination one frame multi-source data after receiving audio, video data, comprise video data, voice data and other various sensors are reported to the police and are identified, and put into FrameBuffer and call for communication thread, finish pre-recording of audio frequency and video simultaneously, in order to warning message is provided.

Rpyroelectric infrared detects thread and wireless receiving thread: receive the alarm signal A i of various alarm sensors, wherein Ai can be the rpyroelectric infrared alerting signal, vibration alarming signal, the broken alerting signal of glass, door magnetic alerting signal etc.When receiving alerting signal, corresponding warning sign Fi puts 1, represent that i sensor report to the police, unison counter Ci puts an initial value, otherwise counter Ci successively decreases, and when counter made zero, putting the sign Fi that reports to the police was 0,, delay counter be set here be for the warning message that guarantees various sensors can be synchronous.

From processor end communication thread: receive the alerting signal of sending from the processor end, notice is reported to the police and is sent thread.If have a frame multi-source data among the FrameBuffer, then these data sent to from the processor end.

Voice output/warning sends thread: after receiving alerting signal, and the warning message that tissue is relevant, as the audio frequency and video of quotation, picture etc. send to exterior terminal, and start equipment alarm such as audible-visual annunciator.If receive voice output information, the opening voice output function connects remote terminal immediately, finishes voice output.

As shown in Figure 5, Intelligent treatment software flow: after at first primary processor end master treatment progress starts, multitask kernel and Intelligent treatment process are loaded into from processor memory, finish a series of initialization from processor, create communication thread automatically and analyze thread.Communication thread at first judges whether to need to send alerting signal, if then send alerting signal to main treatment progress, otherwise judge whether to receive a frame multi-source information, wherein multi-source information comprises: the warning message of audio, video data and other sensor, multi-source information is put into frame buffer FrameBuffer, send the confirmation of receipt signal to main treatment progress simultaneously, show that the Intelligent treatment process is working properly.Analyze thread and from FrameBuffer, read frame data, therefrom take out audio, video data analysis, draw analysis result, in conjunction with other warning message in the Frame, finish multi-source data fusion and false-alarm and filter, determine whether to need to send alerting signal.

Another embodiment of the indoor unexpected abnormality affair alarm of Fig. 6 the present invention system, the judgement of householder's identity comprises by recognition of face carries out householder's identity validation, and wherein said recognition of face mainly is divided into:

The front face photo of same individual's different angles is gathered in the registration of people's face, finds the position of people's face in image by people's face location, and standardization people face size, people's face angle and illumination, the feature of extraction standardization facial image, typing registered user database;

Recognition of face, adopt the HAAR feature to realize that in conjunction with the adaboost algorithm people's face detects realization people face location, adopted two-layer human eye steady arm, all be to obtain the human eye location by the adaboost algorithm, utilize people's face positioning result by the image rotation eyes to be proofreaied and correct and be the standardization of level realization people face, carry out feature extraction by two-dimensional Gabor filter, utilize the covariance between vector to mate the tolerance mode apart from conduct, realize the facial image of inquiry and the coupling of database facial image by the arest neighbors classification.Wherein the Weak Classifier that the Adaboost algorithm is general with a large amount of classification capacities is combined as a strong classifier according to the mode that the training error index descends.And HAAR is characterized as the weak typing feature that the adaboost algorithm provides magnanimity, has guaranteed that the adaboost algorithm totally finds the weak typing of excellent performance.Detect in the implementation process at people's face, the use of integration histogram and cascade classifier greatly reduces the processing time when guaranteeing than high measurement accuracy; Human eye generally is divided into two-layer location during the location, and wherein ground floor is coarse positioning, and locating area has selected to comprise most of ocular of eyes eyebrows, and the second layer is accurately to locate, and locating area only comprises ocular.The coarse positioning device than precise localizer owing to comprised area information more, therefore the stability of location is higher, substantially do not have bigger position deviation, and precise localizer can realize the accurate location of human eye, but the interference that is subjected to eyebrow, canthus easily causes the location mistake.On the basis of coarse positioning, determine the approximate location scope of human eye by the geometric proportion relation, in this scope, use precise localizer to realize the accurate location of human eye.By having reduced eyebrow, canthus etc. to location influence by thick to smart locator meams, improved the accuracy of location; And the standardization of people's face is very crucial in a recognition of face step, and standardization result's quality has directly influenced the precision of recognition of face.Geometry correction and the gamma correction of facial image mainly finished in the standardization of people's face.The result who utilizes the previous step human eye to locate is easy to realize the geometry correction of facial image, at first by the image rotation eyes is proofreaied and correct to be level, intercepts apart from the facial image of d by eyes.As shown in Figure 7, wherein scale the images to the 80x80 pixel at last.

Gamma correction mainly is to eliminate uneven illumination to a certain extent to the influence of follow-up identification.Mainly comprise the plane of illumination match, the deduction plane of illumination, histogram equalization and gray-scale value normalize to zero-mean, unit variance.Here suppose that plane of illumination is a plane.Point on the plane of illumination satisfies following formula: IP (x, y)=to be write as matrix form be x=Np to ax+by+c, the column vector lined up of the picture element gray-scale value of x presentation video wherein, N represents the coordinate of picture element correspondence, horizontal ordinate is shown in first tabulation, secondary series is represented ordinate, and the 3rd row are filled 1, p=[a b c] ^TPlane parameter a, b, c can try to achieve by the mode of linear regression, i.e. p=(N ^TN) ^-1N ^TX.

Can select the Gabor wavelet character, the Gabor conversion has excellent performance aspect the analysis image regional area texture.Two-dimensional Gabor filter

Can be expressed as:

$\mathrm{\&psi;}\left(\stackrel{\&RightArrow;}{x}\right)=\frac{\left|\right|\stackrel{\&RightArrow;}{k}\left|\right|}{{\mathrm{\&sigma;}}^2}\exp (-\frac{{\left|\right|\stackrel{\&RightArrow;}{k}\left|\right|}^2{\left|\right|\stackrel{\&RightArrow;}{x}\left|\right|}^2}{2{\mathrm{\&sigma;}}^2})[\exp \left(i\stackrel{\&RightArrow;}{k}\stackrel{\&RightArrow;}{x}\right)-\exp (-\frac{{\mathrm{\&sigma;}}^2}{2})]$

Wherein

Be image coordinate, Be the centre frequency of wave filter, k _xAnd k _yExpression respectively

In the projection of transverse axis and the longitudinal axis,

Be the direction of wave filter, u and the different value of v representative,

Be Gaussian envelope,

Be the complex values plane wave.Two-dimensional Gabor filter realizes by the sinusoidal wave plane of two-dimensional Gaussian function modulation characteristic frequency and direction, the analysis that the frequency by changing sinusoidal wave plane and direction realize different scale and different directions image texture.

Obtained the facial image of 80x80 size by the standardization of people's face, here 5 wave filter yardsticks have been selected, 8 filtering directions, obtain the Gabor wave filter of 40 different directions and frequency, a facial image is obtained 40 magnitude image behind the Gabor wavelet transformation after by the wave filter convolution, and the Gabor intrinsic dimensionality that obtains at last is 163840.Can reduce the speed of discriminator in the proper vector of such higher-dimension greatly, therefore need carry out dimensionality reduction to proper vector.Here adopt 4x4 evenly to realize the feature dimensionality reduction to down-sampling.

Adopt the arest neighbors classification to realize the facial image of inquiry and the coupling of database facial image, the covariance between the employing vector simultaneously can be by covariance apart from weighing the confidence level of finally mating apart from conduct coupling tolerance mode.

Another embodiment of indoor unexpected abnormality affair alarm system according to the present invention, signal processing module is connected with memory module (Storage), at different applicable cases, can use dissimilar storage mediums, SD and TF cartoon are crossed SDIO and are controlled, and the storage medium of this type is convenient for changing; Nand controls by Nandflash, this type integrated level height, but be not easy to change storage medium; SSD is by PCI-E or the control of SATA interface, and this type storage space can be accomplished very big;

Be provided with Ethernet interface (RJ45);

Be connected with external power supply (DC IN) and battery (Battery) by power management module (Power manager);

Be connected with the internal memory (DDR) for the master cpu operation;

Be connected with the flash memory (Flash) for storage system start-up routine, configuration parameter, log information;

Be connected with wireless module (3G, WIFI), be used for the transmission of audio, video data and remote control signal;

Be connected with wireless module (zigbee, Blue tooth or other wireless modules), be used for receiving the alerting signal that the distributed alarming device sends, as the broken signal of glass and door magnetic opening signal, also can send control information to miscellaneous equipment.

Be connected with by external warning lamp, the warning signal interlocking equipment is realized local sound and light alarm, or controls the interlink warning module of other associate device.

Fig. 8 and Fig. 9 show a preferred embodiment of the indoor unexpected abnormality affair alarm of the present invention system, video acquisition module comprises common shooting detection module and the degree of depth shooting detection module that view data is alignd mutually, obtain sextuple information component＜x for 1 P in the image, y, d, r, g, b＞, wherein＜x, y＞be the coordinate of picture element in image, d is that shot object is apart from the distance of video camera,＜r, g, b＞be color component, this indoor unexpected abnormality affair alarm system for detection of moving target and tracking method as follows:

Coloured image is converted to gray level image, sets up background model at gray level image and depth image respectively, adopt mixed Gaussian to set up background model, the probability distribution that mixed Gaussian background modeling hypothesis signal changes can be used K Gaussian distribution match, is expressed as

$\mathrm{prob}\left(x\right)=\underset{i=1}{\overset{K}{\mathrm{\&Sigma;}}}{\mathrm{\&omega;}}_i{\mathrm{\&eta;}}_i(x,{\mathrm{\&mu;}}_i,{\mathrm{\&sigma;}}_i)$

Wherein μ and σ are average and the variance of Gaussian distribution, each Gaussian distribution η in the model _i(x, μ _i, σ _i) all give a weights omega _i, μ wherein _iAnd σ _iAverage and the standard deviation of representing Gaussian distribution respectively, a plurality of Gaussian distribution obtain the probability distribution of signal by linear combination, K Gaussian distribution is according to the descending sort of ω/σ, arrange forward Gaussian distribution and enough represent the distribution of background, mixed Gauss model can be safeguarded the variation of scene automatically simultaneously, situation about surveying for flase drop can be righted the wrong by study simultaneously

With preceding B Gaussian distribution model as a setting, remaining Gaussian distribution is thought prospect, and B should satisfy

$B=\arg {\min }_b(\underset{n=1}{\overset{b}{\mathrm{\&Sigma;}}}{\mathrm{\&omega;}}_n\left(x\right)>T)$

Present picture element value I (x) and a preceding B Gaussian distribution are mated, if the match is successful then this pixel is background pixels with wherein any one Gaussian distribution, otherwise are the foreground moving pixel, matching way as shown in the formula

|I(x)-μ _i(x)|＜c×σ _i(x)i＝1.....B；

Goal verification, mode by background modeling detects and has obtained moving object, but a lot of motions all are to be changed by indoor light, the swing of curtain, the generation of motion such as the flicker of television image, and the monitored object here is the people, therefore people's motion need be separated from a large amount of irrelevant motions.Here by the light interference filter, people's face detects filtrator and a shoulder detects filtrator, and it is main adopting the gray level image analysis, and the depth image analysis is auxilliary, detects the affirmation that realizes monitoring objective in conjunction with the detection of people's face and head shoulder simultaneously,

Light interference filter wherein: the variation of light makes the gray-scale value of image that variation take place, the accommodation that has surpassed background model when speed and the amplitude of gray-value variation, then detect and be moving target, filter light by depth image and change the motion that causes, depth image is owing to adopted the infrared light detecting method, substantially be not subjected to the interference of illumination variation, detect by the gray level image background modeling and obtain the moving region and be

Zone in the corresponding depth image is

Following condition must be satisfied in real target area:

$\frac{\underset{x\&Element;R_d^i}{\mathrm{\&Sigma;}}{P}_d\left(x\right)}{\underset{x\&Element;R_g^i}{\mathrm{\&Sigma;}}{P}_g\left(x\right)}>T$

P wherein _d(x) and P _g(x) be respectively depth image and gray level image motion detection result, P (x)=1 represents motion pixel, and P (x)=0 represents background pixels;

People's face and head shoulder detect filtrator: have a large amount of real motions in the actual monitored scene, rotation as fan, the swing of curtain, the flicker of television image, the motion of pet, and the present invention only is concerned about people's motion, the people has the obvious external appearance characteristic that is different from these motions, as face characteristic and head shoulder feature, here adopt the haar small echo to realize that in conjunction with the adaboost sorter people's face detects in gray level image, the HOG feature realizes that in conjunction with the svm classifier device head shoulder detects, and is illustrated as the people of motion when detecting face characteristic or head shoulder feature in the moving region, otherwise think to disturb, give filtering;

Target following by the target association of interframe, forms the movement locus of target, for the succeeding target behavioural analysis provides foundation on the basis of goal verification.Target following mainly comprises: target prodiction, and target signature is selected, and target association coupling and target signature are upgraded, wherein

Target prodiction be according to present frame and before the location estimation target of target in the position of next frame, be conducive to improve the precision of subsequent association coupling.Here adopt the mode of estimating target speed, utilize following formula predicted position,

$v_t^x=\frac{\underset{n=0}{\overset{N-2}{\mathrm{\&Sigma;}}}(x_{t-n}-x_{t-n-1})}{N-1}$ With $v_t^y=\frac{\underset{n=0}{\overset{N-2}{\mathrm{\&Sigma;}}}(y_{t-n}-y_{t-n-1})}{N-1}$

Wherein

With

Target is in the speed of x direction and y direction constantly to represent t respectively, and N is time window, x _T-nAnd y _T-nRepresent t-n horizontal ordinate and the ordinate at the extraneous rectangle frame of target center, in like manner x constantly respectively _T-n-1And y _T-n-1Represent t-n-1 horizontal ordinate and the ordinate at the extraneous rectangle frame of target center constantly respectively;

It is two features of center C of selecting relatively stable reliable target area S and target boundary rectangle frame that target signature is selected.The target association coupling is to find the former frame target in the position of present frame, realize the location of target, according to the correlation tracking principle, as long as guarantee under the situation of enough image sampling rates, the change in location of same target between adjacent two frames is not too large, therefore the hunting zone of target can be limited to a less distance range, the while also greatly reduces the risk of mistake coupling, establishes t moment target j to be

T-1 target i constantly is

When satisfying following formula, just carry out characteristic matching:,

$\mathrm{dist}\left(O_t^j\right\{C\},O_{t-1}^i\{C\left\}\right)<\mathrm{\&gamma;}$

Wherein γ is the detection range threshold value, need determine according to actual scene.

The criterion of coupling is following formula

$\arg \min (\mathrm{\&omega;}\&times;\left(\frac{\mathrm{dist}\left(O_t^j\right\{C\},O_{t-1}^i\{C\left\}\right)}{\mathrm{\&gamma;}}\right)+(1-\mathrm{\&omega;})\&times;\frac{\left|O_t^j\right\{S\}-O_{t-1}^i\{S\left\}\right|}{O_{t-1}^i\left\{S\right\}})<T$

Wherein ω is the weight factor of feature, value 0.5, and T is the error upper limit, prevents the mistake coupling, value 0.4.

The data of filtering sudden change are all adopted in the renewal of target signature, guarantee that the single order smooth mode renewal of bigger fluctuation can not appear in data, and specific implementation is the formula of delegation as follows

$O_t^i\left\{F\right\}=\mathrm{\&alpha;}\&times;O_{t-1}^i\left\{F\right\}+(1-\mathrm{\&alpha;})\&times;O_t^i\left\{F\right\},F=\{S,C\}$

Wherein α is for upgrading the factor, value 0.2.

The indoor unexpected abnormality affair alarm of basis system is as follows for detection of the method for falling, the automatic three-dimensional modeling in ground, the video camera that will link to each other with video acquisition module is tilted to down and the angled installation in ground, the initial point of camera coordinate system is set on the Z axle of image coordinate system, and translation matrix T is reduced to like this $\left[\begin{array}{c}0\\ 0\\ H\end{array}\right],$ Wherein H represents the video camera photocentre to the distance on ground, and the conversion formula of camera coordinate system and world coordinate system is as follows

$\left[\begin{array}{c}x\\ y\\ z\\ 1\end{array}\right]=\left[\begin{array}{cc}R& T\\ 0^T& 1\end{array}\right]\left[\begin{array}{c}{x}_w\\ y_w\\ z_w\\ 1\end{array}\right],T=\left[\begin{array}{c}0\\ 0\\ H\end{array}\right],R=R_x\left(\mathrm{\&alpha;}\right)\&CenterDot;R_y\left(\mathrm{\&beta;}\right)\&CenterDot;R_z\left(\mathrm{\&gamma;}\right),$

Three coordinate axis of camera coordinate system and three coordinate axis angles of world coordinate system are α, and β and γ suppose that the initial point of image coordinate system overlaps with the initial point of camera coordinate system, then in the depth image a bit (d) coordinate under camera coordinate system is for u, v

Wherein (u v) is the coordinate of picture element in image, and d is that shot object is apart from the distance of video camera, f _xAnd f _yFocal length for video camera level and vertical direction on imaging plane;

If ground some F (x, y z) satisfy following plane restriction under camera coordinate system:

ax+by+cz+d＝0

Wherein

Be the normal vector of ground level, can calculate the angle α of three coordinate axis, β and γ by normal vector:

$\mathrm{\&alpha;}=\arccos \frac{a}{\left|\stackrel{\&RightArrow;}{v}\right|},$ $\mathrm{\&beta;}=\arccos \frac{b}{\left|\stackrel{\&RightArrow;}{v}\right|},$ $\mathrm{\&gamma;}=\arccos \frac{c}{\left|\stackrel{\&RightArrow;}{v}\right|}$

After obtaining frame depth image data, adopt the RANSAC algorithm to realize plane fitting, can obtain a plurality of candidates' ground level F by match _I=1...n{ a _i, b _i, c _i, d _i, by following priori coarse sizing is carried out on the plane; : one, ground has occupied bigger area in image, and namely real ground level should comprise more image pixel point; Two, generally the angle γ of video camera and z axle between 40 degree and 80 degree, and the angle α of x axle at 0 degree between 20 degree, equal 0 degree substantially with the angle β of y axle.

Select in the remaining plane behind the coarse sizing from camera plane farthest namely to satisfy as ground level:

$F=\underset{1\&le;i\&le;n}{\arg \max }{d}_i$

Any point to the distance of ground level is under the camera coordinates:

$h=\frac{|\mathrm{ax}+\mathrm{by}+\mathrm{cz}+d|}{\sqrt{a^2+b^2+c^2}}$

Calculate target's center to the distance of ground level by following formula;

The static judgement of target, judge by the time-domain difference that calculates gray level image in the target area whether target is static:

$M=\frac{\underset{(x,y)\&Element;R}{\mathrm{\&Sigma;}}|g_t(x,y)-g_{t-1}(x,y)|}{S_R}$

G wherein _t(x, y) constantly in the gray level image (x y) locates the gray-scale value of picture element to expression t, and R represents the target area, S _RElemental area for the target area.The explanation target is static when M＜ε, and wherein ε is a very little positive number.

Generally speaking namely be, at first carry out the ground three-dimensional modeling, determine the ground region in the image, obtain the three-dimensional coordinate on ground, on the basis of target following, calculate the center of target and the distance H of ground level, if H is less than the threshold value T that arranges, people's health close proximity to ground is described, if next detecting the static time of target greater than certain threshold value, then triggers the warning of falling.

Figure 10 shows another preferred embodiment of the indoor unexpected abnormality affair alarm of the present invention system, and the audio collection module also detects specific sound, and its detection method bag is as follows:

The voice signal pre-service, the sampling rate of supposing sound signal X (t) is f _s, f _sValue 8kHz passes through pre-emphasis successively with X (t), divides frame and windowing process, and window function is selected Hanning window, and removes average, avoids DC component that near the spectral line ω=0 place is exerted an influence;

The period map method in the Classical Spectrum estimation is adopted in feature extraction, uses FFT to realize, finally obtains normalized power spectrum X (f _n), the extraction number is 24 Mel bank of filters.Power spectrum X (f _n) through the Mel bank of filters filtering take the logarithm, obtain Mel cepstrum MFCC coefficient through discrete cosine transform again, the Mel bank of filters is made up of one group of V-belt bandpass filter that distributes according to the Mel frequency marking;

The training of GMM audio frequency identification model adopts the maximum EM algorithm of expectation to ask for the GMM model, given training sample set X={x ₁, x ₂..., x _n, the likelihood function of GMM is

$p(X/\mathrm{\&lambda;})=\underset{i=1}{\overset{n}{\mathrm{\&Pi;}}}p(x_i/\mathrm{\&lambda;})$

Model parameter wherein

p _iThe probability of expression Gauss model,

Σ _iMean vector and the covariance matrix of representing Gauss model respectively.

The EM algorithm comprised for two steps, and the E step is asked for expectation, calculates auxiliary function

M goes on foot expectation maximization, maximization

Obtain

The continuous iteration that goes on foot by E step and M is until algorithm convergence,

$Q(\mathrm{\&lambda;},\widehat{\mathrm{\&lambda;}})=\underset{y}{\mathrm{\&Sigma;}}\left(P(Y=y|X=x\left|\mathrm{\&lambda;}\right)\log P(Y=y,X=x|\widehat{\mathrm{\&lambda;}})\right)$

Wherein X is observed reading, and Y is implicit state.

When the expectation value maximal value of adjacent twice iterative computation was more or less the same, then algorithm convergence stopped iteration, is shown below:

$Q_t(\mathrm{\&lambda;},\widehat{\mathrm{\&lambda;}})-Q_t(\mathrm{\&lambda;},\widehat{\mathrm{\&lambda;}})<\mathrm{\&epsiv;}$

Wherein t represents number of iterations, and ε is a less positive number;

The identification of use training pattern obtains model parameter λ by the GMM training, sends in the GMM model after the sound bite extraction feature and calculates similarity, judges the classification of this sound bite by similarity.

The embodiment of indoor unexpected abnormality affair alarm system according to the present invention, a kind of indoor unexpected abnormality affair alarm system comprises light stream vector analysis and audio analysis for detection of the method for limbs conflict, when the conflict of outburst limbs, be accompanied by violent random motion and loud uttering long and high-pitched sounds, therefore can detect the limbs conflict by light stream vector analysis and audio analysis, when two kinds of methods all detect the limbs conflict, trigger the limbs conflict and report to the police, capture a scene photograph simultaneously.

Light stream vector is analyzed, and can obtain the zone at target place by target following, adopts the light stream vector V={ ν in the KLT unique point optical flow computation target area ₁, ν ₂..., ν _n, adopt amplitude weighting histogram H _p={ h _j} _{J=1,2 ..., n}Realize the statistical study of area light flow vector, draw j rank Nogata h by following formula again _j,

$h_j=C_h\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}{A}_{v_i}\mathrm{\&delta;}(b\left(v_i\right)-j)$

Here exponent number can value 12, C _hBe normalized parameter, Be the normalization light stream vector

Amplitude, b (v _i) be light stream vector v _iCorresponding histogram determines that by the direction of vector δ (.) is Kronecker delta function,

Adopt regional entropy E _HRealize the tolerance of violent random motion, E _HExpression formula as follows:

$E_H=-\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{h}_j\log h_j$

H wherein _jRepresent j rank amplitude weighting histogram.E _HMotion Shaoxing opera in the more big declare area is strong random, and setting threshold T works as E _HBroken out the limbs conflict during＞T in the declare area.

Whole workflow of the present invention:

The user can open or close warning function as required, the mode that detects can select video detection, audio detection, rpyroelectric infrared to detect, vibration detection, glass is broken to be detected, the smog inspection detects, during door magnetic detects one or more, as fall, serious attitude conflict adopts audio frequency and video to detect, window area adopts rpyroelectric infrared detection, vibration detection, glass is broken detects, and the door region adopts video detection, audio detection, rpyroelectric infrared detection, vibration detection, door magnetic to detect.Warning message can be video, one or more in audio frequency or the picture.

(1) device power or reset after, signal processing module is load operation system and application program from FLASH, finishes the initialization of main process chip and the configuration of peripheral hardware, next finishes the initialization to each subsystem, enters normal operating conditions at last.When using first, kinsfolk people's face is registered.

(2) audio-video signal and the heat at the continuous acquisition monitoring of the primary processor end scene of main process chip are released infrared detection signal, receive the alerting signal of other distributed alarming sensor simultaneously by wireless receiving module, multi-source data sent into from the processor end analyze, carry out audio frequency and video simultaneously and pre-record.If receive warning or early warning signal from the processor end, then with the audio frequency and video of pre-recording, send to by the alerting signal sending module on cell monitoring center or householder's the mobile phone together with the photo of capturing, start interlocking equipment warnings such as audible-visual annunciator simultaneously.

(3) main process chip moves intelligent audio frequency and video analytical algorithm from the processor end, respectively from the angle analysis of video and audio frequency draw whether exist fall the long period not, serious limbs conflict, stranger unexpected abnormality event such as illegally enter the room takes place, and in conjunction with other sensor warning message and, use the decision-making integration technology to obtain final court verdict, the result is sent to the primary processor end.

(4) after warning is received by householder or Surveillance center, can confirm by the warning message of sending back, but also opening voice intercommunication function, and with the indoor occupant conversation, a situation arises further to understand anomalous event.

The product experiment

Experimental situation and equipment

Experiment place: 120 square metre of 2 Room 2 Room family expenses dwelling house.

Experimental facilities: 1 in the video camera of the interior unexpected abnormality affair alarm system of (1) compartment; (2) the wireless door magnetic detecting device is 1; (3) wireless glass break detector is 2; (4) wireless smoke detector is 1; (5) wireless gas leak sensor is 1; (6) mobile phone of installation unexpected abnormality affair alarm reception process software is 1 one.

As shown in figure 11, ceiling type unexpected abnormality affair alarm video camera is installed in the parlor, respectively in the bedroom, the kitchen, enter front door glass break detector, gas leakage detector, smoke transducer, door magnetic detector be installed, and by wireless and warning camera communication, connect the warning warning signal by warning video camera IO delivery outlet.

Test method and result:

Conclusion (of pressure testing): under indoor environment, test findings is consistent with the test expection, and function, performance meet the product design requirement fully.

The present invention has the following advantages:

Advantage one: function is strong, and purposes is wide.Nearly all unusual accident in the product energy sensing chamber, especially high-precision personnel's fall detection can be widely used in the monitoring to old solitary people.

Advantage two: rate of false alarm is low.The false-alarm filter algorithm that build-in function is powerful, to indoor environment light change, the anomalous event of initiation such as curtain waves, pet carried out filtration preferably.

Advantage three: report to the police in time, warning message is directly perceived, flexible, can be audio frequency, video or picture.

Advantage four: equipment use, installation, easy to maintenance.Adopt wireless transmission, wire laying mode is simple, be easy to working service.

Advantage five: but the free of discontinuities use even have a power failure, still can be opened automatically and carry accumulator continuation use.

Although invention has been described with reference to a plurality of explanatory embodiment of the present invention here, but, should be appreciated that those skilled in the art can design a lot of other modification and embodiments, these are revised and embodiment will drop within the disclosed principle scope and spirit of the application.More particularly, in the scope of, accompanying drawing open in the application and claim, can carry out multiple modification and improvement to building block and/or the layout of subject combination layout.Except modification that building block and/or layout are carried out with improving, to those skilled in the art, other purposes also will be tangible.

Claims (9)

1. an indoor unexpected abnormality affair alarm system is characterized in that: comprise

Video acquisition module adopts video sensor to gather video stream signal, finishes the digitizing of picture signal, and the pre-service of picture signal is met the digital video signal of signal processing module requirement and exports to signal processing module;

The audio collection module is finished the digital-to-analog conversion of voice signal, and the sample code of voice signal and filtering are handled, and is met the digital audio and video signals of signal processing module requirement and exports to signal processing module;

Heat discharges the infrared detection module, the difference of the temperature of inducing moving objects and background object, and heat is released the infrared different information that can sense human body temperature and ambient temperature when human body moves, and converts the output of voltage signal backward signal processing module to;

The vibration detection module produces extraneous vibration deformation or is subjected to force information to change voltage signal into, exports to signal processing module;

The broken detection module of glass and door magnetic detection module are transformed into voltage signal by the wired or wireless communication mode with corresponding information and export to signal processing module;

Signal processing module, collect heat and discharge the signal that infrared detection alarm module, vibration detection module, video acquisition module and audio collection module transmit, carry out analysis-by-synthesis, differentiate illegal invasion, violent conflict, fall do not rise for a long time, anomalous event such as gas leak, and send alerting signal to the warning sending module;

Wireless receiving module is used for receiving the vibration detection module, the broken detection module of glass, the alerting signal that the alarm sensor on smoke detector, emergency call button and the door magnetic detection module sends;

The warning sending module adopts the wired or wireless communication mode, is mainly used in receiving alerting signal that signal processing module sends and to sending warning message to householder's mobile phone or cell management center;

The alarm linkage module, interlock sound and light alarm equipment is worked simultaneously;

The voice output module realizes the voice output from RTU (remote terminal unit) such as mobile phone, Surveillance center;

The infrared LED light supplementation lamp module when ambient light illumination is not enough, provides secondary light source to video acquisition module, guarantees that it still can collect effective view data under the environment of low-light (level);

Power management module connects external power supply and accumulator, when the external power supply normal power supply, entire product is used external power supply, when extraneous power cut-off, inner standby battery is enabled automatically, guarantees that above-mentioned each module continues operate as normal after external power interruption.

2. indoor unexpected abnormality affair alarm according to claim 1 system, it is characterized in that: this system is installed in the casing, described video acquisition module, audio collection module, heat discharge infrared detection alarm module, vibration detection module, warning sending module, wireless receiving module, alarm linkage module and warning sending module and are installed in the casing, described signal processing module and power management module and switch lamp control module are installed in the shell middle part, and the infrared LED light supplementation lamp module is arranged between casing outward flange and the middle part.

3. a kind of indoor unexpected abnormality affair alarm according to claim 1 and 2 system, it is characterized in that: kernel processor chip employing in the described signal processing module " primary processor+from processor " the double-core architecture mode, primary processor is mainly finished the collection of audio-video signal, audio/video coding, the rpyroelectric infrared alerting signal is collected, receive the vibration detection module by wireless receiving module, the broken detection module of glass, door magnetic detection module, smoke detector, the alerting signal that emergency call button distributed alarm sensor sends sends the multi-source alerting signal from processor to together with audio, video data; Be mainly used in moving the audio-video intelligent analysis algorithm from processor, judge whether to exist the unexpected abnormality event by video and sound, simultaneously in conjunction with the rpyroelectric infrared alerting signal, the vibration alarming signal, the broken alerting signal of glass, multi-source informations such as door magnetic alerting signal, form final decision signal, take place if determine anomalous event, send alerting signal to the primary processor end, and start the audible-visual annunciator warning.

4. indoor unexpected abnormality affair alarm according to claim 3 system is characterized in that: described householder's identity is judged and is comprised by recognition of face and carries out identity validation that wherein said recognition of face mainly is divided into

The front face photo of same individual's different angles is gathered in the registration of people's face, finds the position of people's face in image by people's face location, and standardization people face size, people's face angle and illumination, the feature of extraction standardization facial image, typing registered user database;

Recognition of face, adopt the HAAR feature to realize that in conjunction with the adaboost algorithm people's face detects realization people face location, adopted two-layer human eye steady arm, all be to obtain the human eye location by the adaboost algorithm, utilize people's face positioning result by the image rotation eyes to be proofreaied and correct and be the standardization of level realization people face, carry out feature extraction by two-dimensional Gabor filter, utilize Euclidean distance between vector as coupling tolerance mode, realize the facial image of inquiry and the coupling of database facial image by the arest neighbors classification.

5. indoor unexpected abnormality affair alarm according to claim 1 system, it is characterized in that: described signal processing module is connected with memory module;

Be connected with external power supply and battery by power management module;

Be connected with the internal memory for the master cpu operation;

Be connected with the flash memory for storage system start-up routine, configuration parameter, log information;

Be connected with by external warning lamp, the warning signal interlocking equipment is realized local sound and light alarm, or controls the interlink warning module of other associate device.

6. indoor unexpected abnormality affair alarm according to claim 1 system, it is characterized in that: described video acquisition module comprises common shooting detection module and the degree of depth shooting detection module that view data is alignd mutually, obtain sextuple information component＜x for 1 P in the image, y, d, r, g, b＞, wherein＜x, y＞be the coordinate of picture element in image, d is that shot object is apart from the distance of video camera,＜r, g, b＞be color component, this indoor unexpected abnormality affair alarm system for detection of moving target and tracking method as follows:

Coloured image is converted to gray level image, sets up background model at gray level image and depth image respectively, adopt mixed Gaussian to set up background model, the probability distribution that mixed Gaussian background modeling hypothesis signal changes can be used K Gaussian distribution match, is expressed as

$\mathrm{prob}\left(x\right)=\underset{i=1}{\overset{K}{\mathrm{\&Sigma;}}}{\mathrm{\&omega;}}_i{\mathrm{\&eta;}}_i(x,{\mathrm{\&mu;}}_i,{\mathrm{\&sigma;}}_i)$

Wherein μ and σ are average and the variance of Gaussian distribution, and each Gaussian distribution is given a weights omega in the model _i, μ wherein _iAnd σ _iAverage and the standard deviation of representing Gaussian distribution respectively, a plurality of Gaussian distribution obtain the probability distribution of signal by linear combination, and K Gaussian distribution arranged the distribution that forward Gaussian distribution enough represents background according to the descending sort of ω/σ; Mixed Gauss model can be safeguarded the variation of scene automatically simultaneously, and situation about surveying for flase drop can be righted the wrong by study simultaneously.

With preceding B Gaussian distribution model as a setting, remaining Gaussian distribution is thought prospect, and B should satisfy

$B=\arg {\min }_b(\underset{n=1}{\overset{b}{\mathrm{\&Sigma;}}}{\mathrm{\&omega;}}_n\left(x\right)>T)$

Present picture element value I (x) and a preceding B Gaussian distribution are mated, if the match is successful then this pixel is background pixels with wherein any one Gaussian distribution, otherwise are the foreground moving pixel, matching way as shown in the formula

|I(x)-μ _i(x)|＜c×σ _i(x)i＝1.....B；

Goal verification, mode by background modeling detects and has obtained moving object, but a lot of motions all are to be changed by indoor light, the swing of curtain, the generation of motion such as the flicker of television image, and the monitored object here is the people, therefore people's motion need be separated from a large amount of irrelevant motions.Here by the light interference filter, people's face detects filtrator and a shoulder detects filtrator, and it is main adopting the gray level image analysis, and the depth image analysis is auxilliary, detects the affirmation that realizes monitoring objective in conjunction with the detection of people's face and head shoulder simultaneously,

Light interference filter wherein: the variation of light makes the gray-scale value of image that variation take place, the accommodation that has surpassed background model when speed and the amplitude of gray-value variation, then detect and be moving target, filter light by depth image and change the motion that causes, depth image is owing to adopted the infrared light detecting method, substantially be not subjected to the interference of illumination variation, detect by the gray level image background modeling and obtain the moving region and be

Moving region in the corresponding depth image is

Following condition must be satisfied in real target area:

$\frac{\underset{x\&Element;R_d^i}{\mathrm{\&Sigma;}}{P}_d\left(x\right)}{\underset{x\&Element;R_g^i}{\mathrm{\&Sigma;}}{P}_g\left(x\right)}>T$

P wherein _d(x) and P _g(x) be respectively depth image and gray level image motion detection result, P (x)=1 represents motion pixel, and P (x)=0 represents background pixels;

People's face and head shoulder detect filtrator: have a large amount of real motions in the actual monitored scene, rotation as fan, the swing of curtain, the flicker of television image, the motion of pet, and the present invention only is concerned about people's motion, and the people has the obvious external appearance characteristic that is different from these motions, as face characteristic and head shoulder feature.Here in gray level image, adopt the haar small echo to realize that in conjunction with the adaboost sorter people's face detects, the HOG feature realizes that in conjunction with the svm classifier device head shoulder detects, when detecting face characteristic or head shoulder feature in the moving region, be illustrated as the people of motion, otherwise think to disturb, give filtering;

Target following by the target association of interframe, forms the movement locus of target, for the succeeding target behavioural analysis provides foundation on the basis of goal verification.Target following mainly comprises: target prodiction, and target signature is selected, and target association coupling and target signature are upgraded, wherein

Target prodiction be according to present frame and before the location estimation target of target in the position of next frame, be conducive to improve the precision of subsequent association coupling.Here adopt the mode of estimating target speed, utilize following formula predicted position,

$v_t^x=\frac{\underset{n=0}{\overset{N-2}{\mathrm{\&Sigma;}}}(x_{t-n}-x_{t-n-1})}{N-1}$ With $v_t^y=\frac{\underset{n=0}{\overset{N-2}{\mathrm{\&Sigma;}}}(y_{t-n}-y_{t-n-1})}{N-1}$

Wherein

With

Target is in the speed of x direction and y direction constantly to represent t respectively, and N is time window, x _T-nAnd y _T-nRepresent t-n horizontal ordinate and the ordinate at the extraneous rectangle frame of target center, in like manner x constantly respectively _T-n-1And y _T-n-1Represent t-n-1 horizontal ordinate and the ordinate at the extraneous rectangle frame of target center constantly respectively;

It is two features of center C of selecting relatively stable reliable target area S and target boundary rectangle frame that target signature is selected.The target association coupling is to find the former frame target in the position of present frame, realizes the location of target.According to the correlation tracking principle, as long as guarantee under the situation of enough image sampling rates, the change in location of same target between adjacent two frames is not too large, therefore the hunting zone of target can be limited to a less distance range, while also greatly reduces the risk of mistake coupling, establishes t moment target j to be

T-1 target i constantly is When satisfying following formula, just carry out characteristic matching,

$\mathrm{dist}\left(O_t^j\right\{C\},O_{t-1}^i\{C\left\}\right)<\mathrm{\&gamma;}$

Wherein γ is the detection range threshold value, need determine according to actual scene.

The criterion of coupling is following formula

$\arg \min (\mathrm{\&omega;}\&times;\left(\frac{\mathrm{dist}\left(O_t^j\right\{C\},O_{t-1}^i\{C\left\}\right)}{\mathrm{\&gamma;}}\right)+(1-\mathrm{\&omega;})\&times;\frac{\left|o_t^j\right\{S\}-O_{t-1}^i\{S\left\}\right|}{O_{t-1}^i\left\{S\right\}})<T$

Wherein ω is the weight factor of feature, value 0.5, and T is the error upper limit, prevents the mistake coupling, value 0.4.

The data of filtering sudden change are all adopted in the renewal of target signature, guarantee that the single order smooth mode renewal of bigger fluctuation can not appear in data, and specific implementation is the formula of delegation as follows

$O_t^i\left\{F\right\}=\mathrm{\&alpha;}\&times;O_{t-1}^i\left\{F\right\}+(1-\mathrm{\&alpha;})\&times;O_t^i\left\{F\right\},F=\{S,C\}$

Wherein α is for upgrading the factor, value 0.2.

7. indoor unexpected abnormality affair alarm according to claim 1 system is characterized in that: this indoor unexpected abnormality affair alarm system is as follows for detection of the method for falling,

The automatic three-dimensional modeling in ground, the video camera that will link to each other with video acquisition module are tilted to down and the angled installation in ground, and the initial point of camera coordinate system is set on the Z axle of image coordinate system, and translation matrix T is reduced to like this $\left[\begin{array}{c}0\\ 0\\ H\end{array}\right],$ Wherein H represents the video camera photocentre to the distance on ground, and the conversion formula of camera coordinate system and world coordinate system is as follows

$\left[\begin{array}{c}x\\ y\\ z\\ 1\end{array}\right]=\left[\begin{array}{cc}R& T\\ 0^T& 1\end{array}\right]\left[\begin{array}{c}{x}_w\\ y_w\\ z_w\\ 1\end{array}\right],T=\left[\begin{array}{c}0\\ 0\\ H\end{array}\right],R=R_x\left(\mathrm{\&alpha;}\right)\&CenterDot;R_y\left(\mathrm{\&beta;}\right)\&CenterDot;R_z\left(\mathrm{\&gamma;}\right),$

Three coordinate axis of camera coordinate system and three coordinate axis angles of world coordinate system are α, and β and γ suppose that the initial point of image coordinate system overlaps with the initial point of camera coordinate system, then in the depth image a bit (d) coordinate under camera coordinate system is for u, v

Wherein (u v) is the coordinate of picture element in image, and d is that shot object is apart from the distance of video camera, f _xAnd f _yFocal length for video camera level and vertical direction on imaging plane;

If ground some F (x, y z) satisfy following plane restriction under camera coordinate system:

ax+by+cz+d＝0

Wherein

Be the normal vector of ground level, can calculate the angle α of three coordinate axis, β and γ by normal vector:

$\mathrm{\&alpha;}=\arccos \frac{a}{\left|\stackrel{\&RightArrow;}{v}\right|},$ $\mathrm{\&beta;}=\arccos \frac{b}{\left|\stackrel{\&RightArrow;}{v}\right|},$ $\mathrm{\&gamma;}=\arccos \frac{c}{\left|\stackrel{\&RightArrow;}{v}\right|}$

After obtaining frame depth image data, adopt the RANSAC algorithm to realize plane fitting, can obtain a plurality of candidates' ground level F by match _I=1...n{ a _i, b _i, c _i, d _i, by following priori coarse sizing is carried out on the plane; : one, ground has occupied bigger area in image, and namely real ground level should comprise more image pixel point; Two, generally the angle γ of video camera and z axle between 40 degree and 80 degree, and the angle α of x axle at 0 degree between 20 degree, equal 0 degree substantially with the angle β of y axle.

Select in the remaining plane behind the coarse sizing from camera plane farthest namely to satisfy as ground level:

$F=\underset{1\&le;i\&le;n}{\arg \max }{d}_i$

Any point to the distance of ground level is under the camera coordinates:

$h=\frac{|\mathrm{ax}+\mathrm{by}+\mathrm{cz}+d|}{\sqrt{a^2+b^2+c^2}}$

Calculate target's center to the distance of ground level by following formula;

The static judgement of target, judge by the time-domain difference that calculates gray level image in the target area whether target is static:

$M=\frac{\underset{(x,y)\&Element;R}{\mathrm{\&Sigma;}}|g_t(x,y)-g_{t-1}(x,y)|}{S_R}$

G wherein _t(x, y) constantly in the gray level image (x y) locates the gray-scale value of picture element to expression t, and R represents the target area, S _RElemental area for the target area.The explanation target is static when M＜ε, and wherein ε is a very little positive number.

8. indoor unexpected abnormality affair alarm according to claim 1 system, it is characterized in that: the audio collection module also detects specific sound, and its detection method bag is as follows,

The voice signal pre-service, the sampling rate of supposing sound signal X (t) is f _s, f _sValue 8kHz passes through pre-emphasis successively with X (t), divides frame and windowing process, and window function is selected Hanning window, and removes average, avoids DC component that near the spectral line ω=0 place is exerted an influence;

The period map method in the Classical Spectrum estimation is adopted in feature extraction, uses FFT to realize, finally obtains normalized power spectrum X (f _n), the extraction number is 24 Mel bank of filters.Power spectrum X (f _n) through the Mel bank of filters filtering take the logarithm, obtain Mel cepstrum MFCC coefficient through discrete cosine transform again; The Mel bank of filters is made up of one group of V-belt bandpass filter that distributes according to the Mel frequency marking

The training of GMM audio frequency identification model adopts the maximum EM algorithm of expectation to ask for the GMM model, given training sample set X={x ₁, x ₂..., x _n, the likelihood function of GMM is

$p(X/\mathrm{\&lambda;})=\underset{i=1}{\overset{n}{\mathrm{\&Pi;}}}p(x_i/\mathrm{\&lambda;})$

Model parameter wherein

p _iThe probability of expression Gauss model, Σ _iMean vector and the covariance matrix of representing Gauss model respectively.

The EM algorithm comprised for two steps, and the E step is asked for expectation, calculates auxiliary function

M goes on foot expectation maximization, maximization

Obtain

The continuous iteration that goes on foot by E step and M is until algorithm convergence,

$Q(\mathrm{\&lambda;},\widehat{\mathrm{\&lambda;}})=\underset{y}{\mathrm{\&Sigma;}}\left(P(Y=y|X=x\left|\mathrm{\&lambda;}\right)\log P(Y=y,X=x|\widehat{\mathrm{\&lambda;}})\right)$

Wherein X is observed reading, and Y is implicit state.

When the expectation value maximal value of adjacent twice iterative computation was more or less the same, then algorithm convergence stopped iteration, is shown below:

$Q_t(\mathrm{\&lambda;},\widehat{\mathrm{\&lambda;}})-Q_t(\mathrm{\&lambda;},\widehat{\mathrm{\&lambda;}})<\mathrm{\&epsiv;}$

Wherein t represents number of iterations, and ε is a less positive number;

The identification of use training pattern obtains model parameter λ by the GMM training, sends in the GMM model after the sound bite extraction feature and calculates similarity, judges the classification of this sound bite by similarity.

9. indoor unexpected abnormality affair alarm according to claim 1 system, the method that it is characterized in that detecting the limbs conflict is as follows: comprise light stream vector analysis and audio analysis

When the conflict of outburst limbs, be accompanied by violent random motion and loud uttering long and high-pitched sounds, therefore can detect the limbs conflict by light stream vector analysis and audio analysis, when two kinds of methods all detect the limbs conflict, trigger the limbs conflict and report to the police, capture a scene photograph simultaneously.

Light stream vector is analyzed, and can obtain the zone at target place by target following, adopts the light stream vector V={ ν in the KLT unique point optical flow computation target area ₁, ν ₂..., ν _n, adopt amplitude weighting histogram H _p={ h _j} _{J=1,2 ..., n}Realize the statistical study of area light flow vector, draw j rank Nogata h by following formula again _j,

$h_j=C_h\underset{i=1}{\overset{k}{\mathrm{\&Sigma;}}}{A}_{v_i}\mathrm{\&delta;}(b\left(v_i\right)-j)$

Exponent number value 12, C _hBe normalized parameter,

Be the normalization light stream vector

Amplitude, b (v _i) be light stream vector v _iCorresponding histogram determines that by the direction of vector δ (.) is Kronecker delta function, adopts regional entropy E _HRealize the tolerance of violent random motion, E _HExpression formula as follows:

$E_H=-\underset{j=1}{\overset{n}{\mathrm{\&Sigma;}}}{h}_j\log h_j$

H wherein _jRepresent j rank amplitude weighting histogram.E _HMotion Shaoxing opera in the more big declare area is strong random, and setting threshold T works as E _HBroken out the limbs conflict during＞T in the declare area.

Images